Does not respect JULIA_CPU_TARGET

[jakobnissen]

MWE:

$ export JULIA_CPU_TARGET=generic
$ julia --startup-file=no -q

julia> using HostCPUFeatures

julia> has_feature(Val{:x86_64_avx2}())
static(true)

This cause issues during precompilation with pkgimages on Julia 1.9, e.g. BioJulia/FASTX.jl#101.

Without knowing many of the details, I think this should be relatively easy to fix. Setting JULIA_CPU_FEATURES forces precompilation of all packages, so HostCPUFeatures can simply check at precompile time if the env variable is set, and then adjust its compile time results accordingly.

[chriselrod]

HostCPUFeatures.jl/src/HostCPUFeatures.jl

Lines 49 to 50 in a792201

	target = Base.unsafe_string(Base.JLOptions().cpu_target)
	occursin("native", target) || return make_generic(target)

Is where it is supposed to check for this.

[jakobnissen]

This does not work, because:

$ export JULIA_CPU_TARGET=generic
$ julia --startup-file=no -q --project=.
julia> Base.unsafe_string(Base.JLOptions().cpu_target)
"native"

In other words, even though the pkgimages respect the env var JULIA_CPU_TARGET, it is not reflected in Base.JLOptions()

[jakobnissen]

More information can be found here: https://docs.julialang.org/en/v1.10-dev/manual/environment-variables/#JULIA_CPU_TARGET

[chriselrod]

Ah, okay, that makes sense. Want to prepare a PR?
Having it simply fallback to generic is probably fine.

Long term, of course, the plan is to drop this from the ecosystem because it does not integrate well.

[jakobnissen]

I'll give it a go!

[tz-lom]

I think last PR finally brings improvements in this direction